A communication system operation and maintenance management method and system applying data analysis

Abstract

The application provides a communication system operation and maintenance management method and system applying data analysis, relates to the technical field of data processing, and the method comprises the following steps: step 1, collecting the original data flow of equipment operation data and service data in real time, dividing the original data flow according to a preset time window, and mapping the multidimensional data in each window into point cloud data set in a high-dimensional space; step 2, based on the point cloud data set, extracting spatial distribution characteristics through principal component analysis, and constructing a virtual reference plane representing a state; step 3, drawing an initial quadrilateral grid on the virtual reference plane, and adaptively segmenting the initial grid according to the density gradient of the point cloud data set to form a heterogeneous sub-region. The application realizes real-time and accurate early warning of communication system abnormalities, rapid fault positioning and predictive maintenance based on trend analysis, and effectively improves the intelligent level of operation and maintenance management and the system reliability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a communication system operation and maintenance management method and system applying data analysis. BACKGROUND

[0002] In the operation and maintenance scene of communication equipment, the equipment state is affected by factors such as periodic fluctuation of business load and progressive attenuation of hardware performance, and the data characteristics show a dynamic change trend. At the same time, there is a nonlinear coupling relationship between multidimensional data (such as device temperature anomaly may be related to high processing load and decline of heat dissipation module efficiency), so that the state evaluation needs to comprehensively consider the multi-dimensional correlation characteristics. The current technology has room for improvement in the following aspects:

[0003] The traditional method mostly uses fixed threshold for abnormality judgment, and when dealing with scenes such as dynamic change of equipment load (such as reasonable performance fluctuation in business peak period) and performance degradation caused by hardware aging, the adaptability of the threshold needs to be enhanced. When network traffic suddenly changes or the equipment enters the performance degradation stage, the matching degree of the fixed threshold with the actual operation law may be reduced, which to some extent affects the accuracy of abnormality detection. In actual application, the threshold needs to be calibrated regularly combined with artificial experience, and the operation and maintenance efficiency needs to be optimized.

[0004] The equipment operation data and business data have complex distribution characteristics in high-dimensional space (such as differences in index correlation patterns under different business scenarios), but the traditional method mostly sets monitoring rules independently for single-dimensional data, and the mining of spatial distribution characteristics and coupling relationship between multidimensional data is not sufficient. This analysis mode is difficult to fully capture the potential characteristics of the equipment operation state (such as the correlation between traffic surge and signaling processing anomaly), which leads to that the abnormality detection mostly stays on the surface index monitoring, and the recognition ability of the hidden health risk needs to be improved. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a communication system operation and maintenance management method and system applying data analysis, which maps multidimensional data to a point cloud data set in high-dimensional space, constructs a dynamic threshold model combined with principal component analysis and adaptive grid division, and realizes fine description of the equipment operation state.

[0006] To solve the above technical problems, the technical scheme of the present application is as follows:

[0007] In a first aspect, a communication system operation and maintenance management method applying data analysis is provided, which comprises:

[0008] Step 1, real-time collection of original data streams of equipment operation data and business data, segmentation of the original data streams according to a preset time window, and mapping of multidimensional data in each window to a point cloud data set in high-dimensional space;

[0009] Step 2, based on the point cloud dataset, extract spatial distribution features by principal component analysis, and construct a virtual reference plane representing the state;

[0010] Step 3, draw an initial quadrilateral grid on the virtual reference plane, and adaptively segment the initial grid according to the density gradient of the point cloud dataset to form heterogeneous sub-regions;

[0011] Step 4, map the point cloud dataset to the heterogeneous sub-regions, calculate the distribution entropy value of each sub-region point cloud, and calculate the health adjustment coefficient of each sub-region based on the distribution entropy value;

[0012] Step 5, weight and normalize the original data and health adjustment coefficient to generate a preprocessed dataset, and based on the preprocessed dataset, train a dynamic threshold model of device performance and business data by adaptive modeling algorithm;

[0013] Step 6, input real-time preprocessed data into the dynamic threshold model for matching, trigger an early warning when the data deviates from the threshold, and locate the faulty device by associating the physical mapping relationship of the abnormal point aggregation sub-region; Based on the time series trend of the health adjustment coefficient, combined with the output result of the dynamic threshold model, predict the device failure and generate a set of predictive maintenance instructions.

[0014] Further, real-time acquisition of device operation data and business data original data stream, the original data stream is divided according to the preset time window, and the multi-dimensional data in each window is mapped to the point cloud dataset in high-dimensional space, including:

[0015] Divide the collected original data stream into time windows according to the preset fixed time length;

[0016] For n-dimensional data vector in each time window, convert it to coordinate point in m-dimensional space;

[0017] Aggregate all the coordinate points in m-dimensional space obtained by conversion in each time window to form a point cloud dataset representing the state of the time window.

[0018] Further, based on the point cloud dataset, extract spatial distribution features by principal component analysis, and construct a virtual reference plane representing the state, including:

[0019] Perform principal component analysis on the point cloud dataset to calculate the eigenvectors and corresponding eigenvalues of the covariance matrix;

[0020] Filter the eigenvectors based on the eigenvalue size to determine the basis vectors representing the main spatial distribution features;

[0021] A two-dimensional plane is spanned by the basis vectors, which is used as a virtual reference plane representing the spatial distribution features of the point cloud dataset.

[0022] Furthermore, an initial quadrilateral mesh is defined on the virtual reference plane, and the initial mesh is adaptively segmented according to the density gradient of the point cloud dataset to form heterogeneous sub-regions, including:

[0023] Define an initial quadrilateral grid on the virtual reference plane and calculate the density distribution of the projected points of the point cloud dataset on the virtual reference plane.

[0024] Based on the density distribution of the projection points, the density gradient of each part within the initial quadrilateral grid region is determined, and an adaptive region segmentation operation is performed on the initial quadrilateral grid according to the density gradient.

[0025] Through adaptive region segmentation, the virtual reference plane is divided into grid regions with different point cloud density characteristics. These grid regions are called heterogeneous sub-regions.

[0026] Furthermore, the point cloud dataset is mapped to heterogeneous sub-regions, the distribution entropy value of the point cloud in each sub-region is calculated, and the health adjustment coefficient of each sub-region is calculated based on the distribution entropy value, including:

[0027] The points in the point cloud dataset are mapped to the divided grid regions on the virtual reference plane, so that each point uniquely falls into a corresponding heterogeneous sub-region, forming a point cloud subset corresponding to each heterogeneous sub-region.

[0028] For each heterogeneous subregion, the distribution entropy value of the corresponding point cloud subset is calculated based on the spatial distribution characteristics of the corresponding point cloud subset within the corresponding heterogeneous subregion.

[0029] Based on the distribution entropy value of each heterogeneous sub-region, the initial health weight factor for each heterogeneous sub-region is determined through a preset entropy-health mapping relationship.

[0030] The initial health weighting factors of each heterogeneous sub-region are normalized to generate the final health adjustment coefficient for each heterogeneous sub-region.

[0031] Furthermore, the raw data and health adjustment coefficients are weighted and normalized to generate a preprocessed dataset. Based on this preprocessed dataset, a dynamic threshold model for device performance and business data is trained using an adaptive modeling algorithm, including:

[0032] Obtain the original data points in the original data stream and the final health adjustment coefficient for each heterogeneous sub-region;

[0033] Based on the heterogeneous sub-region to which each original data point belongs, the corresponding final health adjustment coefficient is used to weight the original data points to generate weighted data points.

[0034] normalizing all the weighted data points to form a preprocessed data set;

[0035] dividing a historical data part of the preprocessed data set as training input data, learning a dynamic change rule of the device performance data and the service data in the training input data through an adaptive modeling algorithm to obtain a learning result;

[0036] generating a dynamic threshold rule set based on the learning result, the rule set defining a dynamic threshold interval of the device performance and the service data, and encapsulating the dynamic threshold rule set as a dynamic threshold model.

[0037] Further, inputting real-time preprocessed data into the dynamic threshold model for matching, triggering a pre-warning when the data deviates from the threshold, and associating a physical mapping relationship of a sub-region where abnormal points are concentrated to locate a faulty device; predicting a device fault based on a time sequence change trend of a health degree adjustment coefficient and in combination with an output result of the dynamic threshold model, and generating a predictive maintenance instruction set, including:

[0038] inputting real-time preprocessed data into the dynamic threshold model for matching, triggering a device abnormal pre-warning signal when the data exceeds the dynamic threshold interval;

[0039] extracting abnormal data points associated with the pre-warning signal and a sub-region identifier to which the abnormal data points belong, and locating a target faulty device based on a physical mapping relationship of the sub-region;

[0040] continuously collecting health degree adjustment coefficients of each heterogeneous sub-region to generate a time sequence change sequence, analyzing a change rate and a fluctuation intensity feature of the time sequence, and inputting the change rate and the fluctuation intensity feature into the dynamic threshold model to output a device fault probability;

[0041] generating the predictive maintenance instruction set when the fault probability exceeds a preset threshold.

[0042] In a second aspect, a communication system operation and maintenance management system applying data analysis includes:

[0043] a point cloud generation module configured to collect raw data streams of device operation data and service data in real time, divide the raw data streams according to a preset time window, and map multi-dimensional data in each window into a point cloud data set in a high-dimensional space;

[0044] a plane construction module configured to extract spatial distribution features through principal component analysis based on the point cloud data set, and construct a virtual reference plane representing a state;

[0045] a sub-region generation module configured to demarcate an initial quadrilateral grid on the virtual reference plane, and perform adaptive segmentation on the initial grid according to a density gradient of the point cloud data set to form heterogeneous sub-regions;

[0046] The health degree calculation module is configured to map the point cloud data set to heterogeneous sub-regions, count distribution entropy values of the point clouds in the sub-regions, and calculate health degree adjustment coefficients of the sub-regions based on the distribution entropy values.

[0047] The model training module is configured to perform weighted and normalized processing on the original data and the health degree adjustment coefficients, generate a preprocessed data set, and train a dynamic threshold model of the device performance and the business data based on the preprocessed data set by using an adaptive modeling algorithm.

[0048] The maintenance instruction generation module is configured to input real-time preprocessed data into the dynamic threshold model for matching, trigger a warning when the data deviates from the threshold, and locate a faulty device by associating a physical mapping relationship of a sub-region where abnormal points are concentrated.

[0049] In a third aspect, a computing device includes:

[0050] one or more processors;

[0051] a memory device storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0052] In a fourth aspect, a computer-readable storage medium stores a program, when the program is executed by a processor, the method is implemented.

[0053] The above scheme of the present application at least has the following beneficial effects:

[0054] By time window segmentation and high-dimensional space point cloud mapping, device operation data and business data are converted into structured spatial distribution representation, the correlation and dynamic characteristics of multi-dimensional data are retained, and the information fragmentation problem of single-dimensional independent analysis in traditional methods is solved. By using principal component analysis (PCA) to construct a virtual reference plane, high-dimensional data is compressed to two-dimensional space, more than 90% of key information is retained while dimensionality is reduced, and secondary features and noise interference are eliminated, so that the point cloud distribution focuses on the core state characteristics of device operation (such as load fluctuation and business association), and the feature analysis efficiency is improved. By dynamically segmenting the grid based on the point cloud density gradient, heterogeneous sub-regions of different sizes are formed to avoid resolution deficiency or computational redundancy of uniform grids, and resource allocation is optimized. By quantifying the orderliness of sub-region data distribution through distribution entropy, a health degree adjustment coefficient is generated to dynamically represent the health difference of the device in different scenarios. The higher the entropy value (the more unstable the state), the lower the coefficient, which provides differentiated weights for data weighting and threshold model training, and improves the state evaluation accuracy.

[0055] The original data is weighted and normalized in combination with the health coefficient, a dynamic threshold model is trained, and the problem that the traditional fixed threshold cannot adapt to load fluctuation and equipment aging is solved. The model can automatically learn the performance change rule according to historical data, dynamically adjust the threshold interval, and reduce the false alarm and missed alarm rate. Through the physical mapping of the sub-region, the abnormal data is quickly associated with the actual equipment, the fault positioning is converted from manual troubleshooting to data-driven intelligent mapping, and the response time is shortened; based on the health coefficient time series analysis and dynamic threshold output, the fault probability is predicted and the maintenance instruction is generated, the operation and maintenance are promoted from 'passive response' to 'active prevention', and the risk of sudden failure is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 is a flowchart of a communication system operation and maintenance management method using data analysis provided by an embodiment of the present application.

[0057] Figure 2 is a schematic diagram of a communication system operation and maintenance management system using data analysis provided by an embodiment of the present application. DETAILED DESCRIPTION

[0058] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0059] As Figure 1 shown, an embodiment of the present application proposes a communication system operation and maintenance management method using data analysis, which comprises the following steps:

[0060] Step 1, real-time acquisition of original data stream of equipment operation data and service data, segmentation of the original data stream according to a preset time window, and mapping of multi-dimensional data in each window into point cloud data set in high-dimensional space;

[0061] Step 2, based on the point cloud data set, extracting spatial distribution features by principal component analysis, and constructing a virtual reference plane representing the state;

[0062] Step 3, dividing an initial quadrilateral grid on the virtual reference plane, and performing adaptive segmentation on the initial grid according to the density gradient of the point cloud data set to form heterogeneous sub-regions;

[0063] Step 4, mapping the point cloud data set to the heterogeneous sub-regions, calculating the distribution entropy value of the point cloud of each sub-region, and calculating the health adjustment coefficient of each sub-region based on the distribution entropy value;

[0064] Step 5, the original data and the health adjustment coefficient are weighted and normalized to generate a preprocessed data set, and based on the preprocessed data set, a dynamic threshold model of device performance and business data is trained by an adaptive modeling algorithm;

[0065] Step 6, real-time preprocessed data is input into the dynamic threshold model for matching, a warning is triggered when the data deviates from the threshold, and the physical mapping relationship of the abnormal point aggregation sub-area is associated to locate the faulty device; based on the time sequence change trend of the health adjustment coefficient and the output result of the dynamic threshold model, the device failure is predicted, and a predictive maintenance instruction set is generated.

[0066] In the embodiment of the application, the health adjustment coefficient is generated by quantifying the orderliness of the sub-area data distribution through distribution entropy, which dynamically represents the health state difference of the device under different operating scenarios. The higher the entropy value (the more disordered the data distribution), the more unstable the device state, which provides a differentiated weight factor for data weighting and threshold model training. The original data is weighted and normalized in combination with the health adjustment coefficient, and a dynamic threshold model is trained to solve the problem that the traditional fixed threshold cannot adapt to load fluctuations and device aging. The model can dynamically learn the device performance change law according to historical data, automatically adjust the threshold interval, and improve the accuracy and environmental adaptability of anomaly detection.

[0067] Through the sub-area physical mapping relationship, the actual device corresponding to the abnormal data is quickly associated, the fault location is converted from manual troubleshooting to data-driven intelligent mapping, and the fault response time is shortened; based on the time sequence analysis of the health coefficient and the output of the dynamic threshold, the device failure probability is predicted and the maintenance instruction is generated, which promotes the operation and maintenance mode from "response after failure" to "prevention before failure", reduces the risk of sudden failure and maintenance cost.

[0068] In a preferred embodiment of the application, the above step 1, the original data stream of device operation data and business data is collected in real time, the original data stream is divided according to a preset time window, and the multi-dimensional data in each window is mapped into a point cloud data set in a high-dimensional space, which can include:

[0069] Step 100, the collected original data stream is divided into time windows according to a preset fixed time length;

[0070] Step 101, for the n-dimensional data vector in each time window, it is converted into a coordinate point in an m-dimensional space;

[0071] Step 102, aggregate all the coordinate points in the m-dimensional space obtained by conversion in each time window to form a point cloud data set representing the state of the time window.

[0072] In the embodiments of the present application, real-time access device operation data and service data are generated in the form of continuous flow. The device operation data includes, but is not limited to, real-time monitoring indicators such as device CPU utilization, memory occupancy, port traffic, hardware temperature, and signaling processing delay; the service data covers user concurrency, data throughput, service transaction success rate, service response time, and other key parameters related to service operation. These data are generated in real time in the form of continuous flow, containing time stamp, device identification, indicator dimension, and corresponding value information.

[0073] The original data flow is divided according to a preset fixed time interval (for example, 1 minute, 5 minutes, or a custom length), forming continuous and non-overlapping time windows. Each time window is identified by a start time stamp and an end time stamp, for example, the first window is 00:00-00:05, the second window is 00:05-00:10, and so on. During the division process, invalid records with abnormal time stamps or missing data are automatically filtered out to ensure the integrity of the data in each window.

[0074] In step 101, for the data flow in each time window, all valid data records are first extracted, grouped by device identification or service type (for example, by device type such as base station, switch, server, or by service type such as voice service, data service), and an n-dimensional data vector in the window is generated. Here, "n-dimensional" corresponds to the number of data characteristics, for example, if CPU utilization, memory occupancy, and port traffic are collected as three device indicators and user concurrency and service success rate are collected as two service indicators, then n=5, forming a 5-dimensional data vector. The n-dimensional data vector is mapped to an m-dimensional spatial coordinate point (m=n), for example, the 5-dimensional data (CPU, memory, temperature, traffic, and delay) is mapped to a 5-dimensional spatial point, retaining the spatial correlation between indicators (for example, the positive correlation between CPU and temperature is represented as the clustering trend of the coordinate point in space). During the conversion process, the dimension and value range of the original data are retained (normalization is not performed at this time), and only the dimension mapping is used to convert the multi-dimensional data into spatial coordinates, ensuring that each data point completely carries the original feature information.

[0075] Step 102, after completing the conversion of all n-dimensional data vectors in a single time window to m-dimensional space coordinate points, aggregate all coordinate points in the same window to form a point cloud data set corresponding to the time window. Each point cloud data set contains the state data points of all devices or services in the window, the number of points is equal to the number of valid data records in the window, and the spatial distribution of the points reflects the correlation characteristics of the multi-dimensional data in the time period (for example, whether some indicators increase or decrease at the same time, whether the data points show a clustering or dispersion trend, etc.). Attach metadata information to each point cloud data set, including the start and end time of the time window, the type of device or service, the data collection frequency, etc. The point cloud data sets of all time windows are stored in chronological order to form a dynamically updated point cloud data sequence.

[0076] By dividing the fixed time window, the continuous original data stream is converted into discrete time slices, avoiding feature ambiguity caused by data mixing. Mapping n-dimensional data vectors to m-dimensional space coordinate points directly preserves the correlation between different indicators (for example, the synchronous change of device temperature and CPU utilization can be reflected by the coordinate correlation of space points), solving the information fragmentation problem caused by single-dimensional independent analysis in traditional methods. The point cloud data set formed by aggregation presents the overall characteristics of multi-dimensional data (such as clustering, dispersion, distribution trend, etc.) in the form of geometric spatial distribution, improving the comprehensiveness and effectiveness of data processing.

[0077] In a preferred embodiment of the present application, the above step 2, based on the point cloud data set, extracts spatial distribution characteristics through principal component analysis to construct a virtual reference plane representing the state, which can include:

[0078] Step 200, principal component analysis is performed on the point cloud data set to calculate the eigenvectors and corresponding eigenvalues of the covariance matrix;

[0079] Step 201, based on the size of the eigenvalue, the eigenvectors are filtered to determine the basis vectors representing the main spatial distribution characteristics;

[0080] Step 202, a two-dimensional plane is spanned by the basis vectors as a virtual reference plane representing the spatial distribution characteristics of the point cloud data set.

[0081] In an embodiment of the present application, the m-dimensional point cloud data set generated by the acquisition step 102 is obtained, where each data point represents the coordinates of the device operation and service data in the m-dimensional space (for example, m=5 corresponds to CPU utilization, memory occupancy, etc. 5 dimensions). To eliminate the influence of the absolute position of different dimensions of data, first, centralize all data: calculate the mean value of each dimension, and then subtract the mean value from the value of all data points in that dimension to shift the center of each dimension data to the origin. For example, if the mean value of the CPU utilization dimension is 60%, then subtract 60% from the value of each data point in that dimension to get the relative value after centralization.

[0082] After the centralization of the point cloud data (subtracting the mean value of each dimension from the data to move the data center to the origin), the covariance matrix C is calculated, which is an m x m square matrix (m is the number of data dimensions, such as 5-dimensional data corresponding to a 5 x 5 matrix). The following is an example of calculating the elements Ci, j in the matrix with two specific dimensions (dimension i = "CPU utilization" and dimension j = "device temperature"):

[0083] Assume that the current point cloud data set contains n centralized data points, each with a value of xi, x2,..., xn in dimension i (the mean value has been subtracted, so the mean value is 0), and a value of yi, y2,..., yn in dimension j (the mean value is also 0).

[0084] The calculation method of covariance Ci, j is to multiply each data point's xk and yk to get x1y1, x2y2,..., xnyn, and then take the average of the n products (i.e., the sum divided by n). If the result is positive, it means that when the value of dimension i is higher than the mean value, the value of dimension j also tends to be higher than the mean value (e.g., CPU utilization increases, temperature also increases); if it is negative, it means that the two tend to change in the opposite direction (e.g., memory usage increases, cache hit rate decreases); if it is close to zero, it means that there is no obvious linear correlation between the two.

[0085] Variance calculation (diagonal elements):

[0086] Diagonal element Ci, i is the variance of dimension i, calculated by multiplying each data point's xk 2 and taking the average (i.e., (x1 2 + x2 2 +... + xn 2 ) ÷ n). The larger the variance, the more dispersed the centralized values of the dimension data (e.g., a large temperature variance means that the device temperature fluctuates greatly around the mean value); the smaller the variance, the more concentrated the data (e.g., the voltage fluctuation variance of a stable running device tends to zero). Traverse all dimension pairs (i from 1 to m, j from 1 to m) to fill each element of the covariance matrix, and finally get a symmetric matrix (Ci, j = Cj, i), whose diagonal is the variance of each dimension and the non-diagonal is the covariance of each pair of dimensions.

[0087] After the covariance matrix C is constructed, it is subjected to eigenvalue decomposition, the goal of which is to find a set of orthogonal eigenvectors and their corresponding eigenvalues. The specific steps are as follows:

[0088] The core of eigenvalue decomposition is to solve the equation C x v = λ x v, where v is the eigenvector and λ is the corresponding eigenvalue. This equation represents that the vector v remains unchanged (or reversed) in direction and is scaled by λ times in length after being transformed by the matrix C. Each eigenvector vk represents a "main variation direction" of the data in the m-dimensional space, which is a linear combination of the original dimensions. For example, for 5-dimensional data (CPU, memory, temperature, traffic, latency), the first eigenvector may correspond to a positive correlation combination direction of "CPU + temperature + traffic" (i.e., these three indicators tend to rise or fall at the same time), and the second eigenvector may correspond to a negative correlation combination direction of "memory + latency" (when memory usage increases, latency tends to decrease, reflecting the optimization effect of the cache mechanism).

[0089] The eigenvalue λ represents the variance of the data in the direction of the corresponding eigenvector. The larger λ is, the more dispersed the data is in that direction (the more information it carries); on the contrary, the smaller λ is, the more concentrated the data is (the less information it carries). All eigenvalues are sorted in descending order (e.g., λ1≥ λ2≥... ≥ λm), and each eigenvector is ensured to be a unit vector (length 1) and any two different eigenvectors are orthogonal (dot product is 0), thus forming an orthonormal basis.

[0090] In step 201, all eigenvalues are sorted in descending order, and the corresponding eigenvectors are arranged accordingly. Since a two-dimensional plane needs to be constructed later, the first two largest eigenvalues corresponding to the eigenvectors are selected, denoted as v1 and v2. These two vectors correspond to the two directions with the largest data variation, for example:

[0091] v1 may correspond to a comprehensive change direction related to device load (e.g., a trend of simultaneous increase in CPU, memory, and traffic), with the largest eigenvalue λ1, indicating the highest degree of dispersion of data in this direction;

[0092] v2 may correspond to a secondary change direction related to business fluctuations (e.g., associated changes in user concurrency and response latency), with the second largest eigenvalue λ2, reflecting a less important data distribution characteristic.

[0093] During screening, the cumulative contribution rate of the first two eigenvalues (i.e., the proportion of λ1+λ2 to the sum of all eigenvalues) is evaluated, which is usually required to exceed 80% or 90% to ensure that sufficient original data information is retained. For directions with smaller eigenvalues (e.g., λ3 and beyond), since the data is highly concentrated in these directions (small variance), they are considered as secondary characteristics or noise and can be ignored.

[0094] Step 202, a "virtual reference plane" is constructed with v1 and v2 as the coordinate axes of a two-dimensional coordinate system. The two vectors are orthogonal to each other in the m-dimensional space, so the plane is a two-dimensional subspace. To map the original point cloud data to the plane, perform a projection operation on each m-dimensional data point P: calculate the projection coordinates of P in the v1 and v2 directions, that is, calculate the dot product of P and v1, P and v2 (which can be understood as "the component size of P in the v1 direction"), to get the two-dimensional coordinates (P x v1, P x v2).

[0095] For example, if the m-dimensional coordinates of the original point P are (x1, x2, x3, x4, x5), and the components of v1 are (a1, a2, a3, a4, a5), then the projection coordinates of P in the v1 direction are x1 x a1 + x2 x a2 + … + x5 x a5, and the projection coordinates in the v2 direction can be obtained in the same way. After all the points are projected, a new point cloud distribution is formed on the two-dimensional plane, which retains the most important spatial features of the original data (such as the point cloud aggregation area in the normal state, the dispersion trend in the abnormal state), while eliminating the secondary information interference in the high-dimensional space.

[0096] By principal component analysis, the m-dimensional data is compressed to a two-dimensional plane, and the secondary features and noise interference are removed, so that the point cloud distribution is more focused on the core state features of the device operation (such as load fluctuation, service association, etc.), and the calculation efficiency is improved by about (m-2) ÷ m times. The point cloud distribution on the two-dimensional plane can directly reflect the clustering, dispersion trend and change direction of the data (such as the point cloud concentrating in a certain area in the normal state, and diffusing to the edge in the abnormal state), which is convenient for identifying the device state mode. Different dimensions and different types of data (such as temperature, flow, user concurrency) are converted into two-dimensional coordinates, eliminating the influence of dimension difference on analysis, and making the correlation features of multi-dimensional data (such as the positive correlation between temperature and load) directly reflected by the spatial distance.

[0097] In a preferred embodiment of the present application, step 3 above, the initial quadrilateral grid is divided on the virtual reference plane, and the initial grid is adaptively segmented according to the density gradient of the point cloud data set to form heterogeneous sub-regions, which can include:

[0098] Step 300, an initial quadrilateral grid is divided on the virtual reference plane, and the density distribution of the projection points of the point cloud data set on the virtual reference plane is calculated;

[0099] Step 301, based on the density distribution of the projection points, the density gradient of each part in the initial quadrilateral grid region is determined, and an adaptive regional segmentation operation is performed on the initial quadrilateral grid according to the density gradient;

[0100] Step 302, the virtual reference plane is divided into divided grid regions with different point cloud density characteristics by an adaptive region segmentation operation, and the divided grid regions are heterogeneous sub-regions.

[0101] In the embodiment of the application, all the projection points on the virtual reference plane generated in the acquisition step 202 are obtained, and the coordinate ranges of the points on a two-dimensional plane are determined: the minimum and maximum values of the x-axis and the minimum and maximum values of the y-axis are found, and a minimum axis-aligned quadrilateral (rectangle) containing all the projection points is determined as an initial grid with the four values as boundaries. The rectangle completely covers the point cloud distribution region.

[0102] To quantify the distribution density of the point cloud on the virtual plane, a kernel density estimation (KDE) method (a non-parametric method that does not depend on a preset distribution model) is used. The specific process is as follows:

[0103] For any position (x, y) on the virtual plane, a circular window with a radius of h is set with the point as the center (the bandwidth h can be adaptively adjusted according to the data size, such as 1.5 times the average distance of all points);

[0104] The number of projection points falling into the window is counted, and a weight is assigned according to the distance from the center (such as a Gaussian kernel function, the closer the distance, the higher the weight), and the density value p(x, y) of the position is calculated; traverse the grid nodes (or uniformly sampled points) on the virtual plane to generate a two-dimensional density distribution matrix, and each element in the matrix corresponds to the density value of a position, forming a continuous density distribution surface. Finally, the density distribution result is presented in a color gradient chart: the darker the color (such as red), the higher the density (point cloud aggregation area), and the lighter the color (such as blue), the lower the density (point cloud sparse area).

[0105] Step 301, for any position (x, y) on the virtual reference plane, the point cloud density is obtained by the kernel density estimation method of step 300, denoted as p(x, y) (unit: number of points / unit area), which reflects the aggregation degree of the point cloud near the position. The density gradient is a two-dimensional vector , wherein:

[0106] is the x-direction gradient, which is approximated by the finite difference method , which represents the horizontal density change rate (dx is a very small step, which is 1÷10 of the current grid resolution, such as 0.1 unit);

[0107] is the y-direction gradient, which is approximated by , which represents the vertical density change rate (dy has the same value as dx).

[0108] The modulus of the gradient vector The gradient indicates the degree of density change. The larger the gradient, the faster the point cloud density changes from sparse to dense (or vice versa) near that location (e.g., edge regions). The gradient direction points in the direction of the fastest density increase and is used to identify the aggregation trend of point clouds.

[0109] The recursive quadtree algorithm is used to adaptively divide the initial quadrilateral grid. The specific steps are as follows:

[0110] Using the smallest quadrilateral containing all point clouds as defined in step 300 as the root node, calculate the maximum density gradient value (i.e., the maximum value of the gradient magnitude of all points in the region) and the total number of point clouds within that region.

[0111] Set two split trigger conditions:

[0112] Condition 1: The maximum density gradient value in the region exceeds the preset threshold T, indicating that the density in the region changes drastically (such as the edge of the point cloud or the boundary of different state clusters). Here, T is set as the mean of the density gradient of historical data + 1 standard deviation (e.g., T = 5 points / unit area / unit length).

[0113] Condition 2: If the number of point clouds in the region exceeds the empirical value N (for example, for a medium-sized cluster (100 devices), N = 100 points / region), it indicates that the data in the region is dense and needs to be further subdivided to capture detailed features. If the root node satisfies any of the partitioning conditions, it is divided into four sub-quadrilaterals (top left, top right, bottom left, bottom right), with each sub-region having a side length that is half that of the original region.

[0114] Repeat the following operation for each sub-quadrilateral:

[0115] Calculate the maximum density gradient value and the number of point clouds within the sub-region; if the sub-region simultaneously satisfies "maximum density gradient value ≤ T" and "number of point clouds ≤ N", then stop the segmentation and mark it as a "leaf node" (i.e., the final heterogeneous sub-region); otherwise, continue the recursive segmentation until all sub-regions satisfy the termination condition.

[0116] Differential processing between high-density gradient regions and low-density gradient regions:

[0117] When there are points within a region whose density gradient magnitude exceeds the threshold T, or whose average gradient magnitude is greater than T / 2, it is determined to be a high-density gradient region. Typical scenarios include:

[0118] The edge areas of point cloud distribution (such as the transition zone between normal and abnormal states, where the density drops sharply from high to low values, with a drastic gradient);

[0119] Boundary area of multi-cluster point cloud (such as the transition zone of different operation modes of the device, the density change presents a complex gradient). For such areas, smaller sub-regions are generated by recursive segmentation (such as subdivision from an initial edge length of 20 units to 5 units) to capture the density mutation characteristics with high precision and avoid missing the distribution details of abnormal points.

[0120] Low-density gradient area:

[0121] When the density gradient module length of all points in the area is ≤T, and the average gradient module length is less than T / 2, it is determined as a low-density gradient area, and the typical scenarios include:

[0122] Point cloud core aggregation area (such as the state point aggregation area of the device when it is stably running, the density distribution is uniform, and the gradient is gentle);

[0123] Sparse data edge blank area (far from the main point cloud cluster, the number of point clouds is small and the density change is weak). For such areas, the system maintains a larger grid edge length (such as 10-20 units) to avoid invalid subdivision due to small data volume or stable density and reduce the consumption of computing resources.

[0124] Step 302, after recursive segmentation, the virtual reference plane is divided into multiple quadrilaterals of different sizes and irregular shapes (since the quadtree segmentation always maintains a quadrilateral structure, but the edge length of the sub-regions at different levels is different), these sub-regions have the following characteristics:

[0125] High-density sub-region: small edge length (such as 10×10 units), corresponding to the core area of point cloud aggregation (such as the state point aggregation area of the device when it is normally running), low internal density gradient, and uniform point cloud distribution;

[0126] Low-density sub-region: large edge length (such as 50×50 units), corresponding to the sparse edge area of point cloud (such as the state point dispersion area of the device when it is abnormal), high internal density gradient, and uneven point cloud distribution.

[0127] Each sub-region records the following metadata:

[0128] Geometric information: coordinates of the upper left corner and the lower right corner, edge length, area;

[0129] Density characteristics: average density, maximum density, density gradient range;

[0130] Point cloud statistics: the number of projection points contained, and the device identifier set (associated through the device grouping information of step 100).

[0131] The final heterogeneous sub-regions form a "density-space" hierarchical structure: high-density areas are used to capture subtle state differences of the device when it is normally running, and low-density areas are used to identify the edge distribution of abnormal points.

[0132] Through density gradient-driven adaptive segmentation, the virtual plane is divided into heterogeneous sub-regions reflecting the degree of data aggregation. Fine meshes in high-density regions capture subtle differences in normal conditions, while coarse meshes in low-density regions identify the edge distribution of anomalies. This avoids the shortcomings of traditional uniform meshes in depicting complex data distributions (e.g., insufficient resolution in high-density regions and computational redundancy in low-density regions). Fine-grained segmentation is automatically performed in data-dense regions (e.g., the core state area of ​​normal equipment operation) while maintaining coarse-grained segmentation in sparse regions (e.g., anomaly edge areas), reducing unnecessary computation. The geometric boundaries of the heterogeneous sub-regions directly correspond to the boundary differences in equipment operating states (e.g., high-density sub-regions correspond to "stable operating state clusters," and low-density sub-regions correspond to "abnormal fluctuation state clusters"), upgrading equipment health assessment from a single indicator judgment to a comprehensive analysis of multi-dimensional spatial distribution characteristics. Each sub-region is associated with a specific set of device identifiers (through the device grouping information in step 100). When anomalies are clustered in a low-density sub-region, the corresponding device group or individual device can be located directly through the physical mapping relationship of the sub-region, avoiding the ambiguity of the association between anomalies and devices in traditional methods, and achieving accurate mapping of "spatial location → device entity".

[0133] In a preferred embodiment of the present invention, step 4 above, which maps the point cloud dataset to heterogeneous sub-regions, calculates the distribution entropy value of the point cloud in each sub-region, and calculates the health adjustment coefficient of each sub-region based on the distribution entropy value, may include:

[0134] Step 400: Map the points in the point cloud dataset to the divided grid regions on the virtual reference plane, so that each point uniquely falls into a corresponding heterogeneous sub-region, forming a point cloud subset corresponding to each heterogeneous sub-region;

[0135] Step 401: For each heterogeneous sub-region, calculate the distribution entropy value of the corresponding point cloud subset based on the spatial distribution characteristics of the corresponding point cloud subset within the corresponding heterogeneous sub-region.

[0136] Step 402: Based on the distribution entropy value of each heterogeneous sub-region, determine the initial health weight factor for each heterogeneous sub-region through a preset entropy-health mapping relationship;

[0137] Step 403: Normalize the initial health weight factors of each heterogeneous sub-region to generate the final health adjustment coefficient for each heterogeneous sub-region.

[0138] In this embodiment of the invention, each heterogeneous sub-region has a defined geometric boundary on the virtual reference plane, determined by the quadtree segmentation in step 301, and recorded as the coordinates of the upper left corner ( ) and the coordinates of the lower right corner ( )(assuming the positive direction of the y-axis is downward). For any projected point (x, y) in the point cloud dataset,

[0139] If the following conditions are satisfied, ≤ ≤ and ≤ ≤ , the point falls into the current sub-region; since the sub-regions of the quadtree partition do not overlap and completely cover the virtual plane, each point must belong to only one sub-region. Traverse all the projected points and distribute the points into the corresponding sub-regions according to the above attribution rules to form a point cloud subset corresponding to each heterogeneous sub-region. For example, if the virtual plane is divided into 10 sub-regions, sub-region A contains 30 points and sub-region B contains 15 points, each subset stores the two-dimensional coordinates of all points in the sub-region and the associated metadata such as original device identifier and timestamp.

[0140] Step 401, to quantify the uniformity of the distribution of the point cloud in the sub-region, the continuous space of each heterogeneous sub-region is divided into k x k equidistant grid cells (k can be adaptively set according to the side length of the sub-region, for example, a sub-region with a side length of 20 units is divided into 10 x 10 grid cells with a cell side length of 2 units). The coordinate range of each grid cell is where , = 0, 1,..., k-1.

[0141] For each point cloud subset corresponding to a sub-region, the number of points in each grid cell is counted, and the proportion of the number of points in the cell to the total number of points in the sub-region is calculated = the number of points in each grid cell ÷ the proportion of the number of points in the cell to the total number of points in the sub-region (if there are no points in the cell, then = 0). The distribution entropy value measures the degree of disorder of the point cloud distribution based on the information entropy principle, and the calculation formula is: distribution entropy value where, = 0 corresponds to 0 (since 0 ln(0) is defined as 0). The entropy value characteristics are as follows:

[0142] When all points are concentrated in one grid cell ( = 1 and the rest = 0), the distribution entropy value = 0, indicating that the distribution is completely ordered (stable state);

[0143] When the point cloud is uniformly distributed in all grid cells, the distribution entropy value reaches the maximum value, indicating the most disordered distribution (unstable state).

[0144] ​Step 402, a negative correlation mapping relationship between distribution entropy value and equipment health degree is established through historical data fitting: the higher the entropy value, the more dispersed the point cloud distribution (the equipment state fluctuates greatly), and the lower the health degree; the lower the entropy value, the more concentrated the distribution (the state is stable), and the higher the health degree. The specific mapping steps are as follows:

[0145] Determine the mapping interval: assuming that the distribution entropy value range is [distribution entropy value min, distribution entropy value max] (obtained by entropy value statistics of all sub-regions), and the health degree weight factor range is pre-set as [Wmin, Wmax] (for example, [0.1, 1.0]);

[0146] Initial weight factor , which ensures that high entropy corresponds to low weight and low entropy corresponds to high weight.

[0147] If the point number of a point cloud subset of a sub-region is less than 5 (the data amount is too small), it is considered that the entropy value calculation is unreliable, and the initial weight factor is directly set to a default value (such as 0.5) to avoid noise interference.

[0148] Step 403, in order to make the health degrees of all sub-regions comparable, the initial weight factors of all sub-regions are normalized to ensure that the sum of the final adjustment coefficients is 1. The specific steps are as follows:

[0149] Calculate the sum of the initial weight factors of all sub-regions, and the final health degree adjustment coefficient = initial weight factor ÷ sum of initial weight factors of all sub-regions. The final health degree adjustment coefficient ∈ (0, 1), which represents the health degree of the equipment state corresponding to the point cloud of the sub-region:

[0150] The final health degree adjustment coefficient is close to 1: the point cloud distribution is concentrated (the entropy value is low), the equipment state is stable, and the health degree is high;

[0151] The final health degree adjustment coefficient is close to 0: the point cloud distribution is dispersed (the entropy value is high), the equipment state fluctuates greatly, and the health degree is low.

[0152] The point cloud spatial distribution characteristics are converted into a calculable health degree index through the distribution entropy value, realizing the leap from "qualitative observation" to "quantitative analysis". For example, a stable running equipment corresponds to a low entropy value (the coefficient is close to 1), and an aging or abnormal equipment corresponds to a high entropy value (the coefficient decreases), providing a scientific basis for subsequent data weighting. The normalized health degree adjustment coefficient can adapt to changes in the equipment running scene: automatically reducing the weight of the corresponding sub-region during the business peak period (the point cloud distribution may be more dispersed, and the entropy value increases) to avoid misjudgment; and increasing the weight during the equipment stable period (the point cloud is concentrated, and the entropy value decreases) to enhance the feature expression of the normal state. The time sequence change (such as continuous decrease) of the health degree adjustment coefficient can be used as an early signal of equipment degradation.

[0153] In a preferred embodiment of the present application, the step 5 of weighting and normalizing the original data and the health adjustment coefficient to generate a preprocessed data set, and training a dynamic threshold model of device performance and service data based on the preprocessed data set through an adaptive modeling algorithm, can include:

[0154] Step 500, obtaining original data points in the original data stream and final health adjustment coefficients corresponding to each heterogeneous sub-region;

[0155] Step 501, according to the heterogeneous sub-region to which each original data point belongs, using the corresponding final health adjustment coefficient to perform weighting processing on the corresponding original data point to generate a weighted data point;

[0156] Step 502, normalizing all weighted data points to form a preprocessed data set;

[0157] Step 503, dividing the historical data part of the preprocessed data set as training input data, learning the dynamic change rule of device performance data and service data in the training input data through an adaptive modeling algorithm to obtain a learning result;

[0158] Step 504, based on the learning result, generating a dynamic threshold rule set, the rule set defining the dynamic threshold interval of device performance and service data, and encapsulating the dynamic threshold rule set as a dynamic threshold model.

[0159] In an embodiment of the present application, the original data points extracted from the real-time data stream include device running data (such as CPU utilization, memory occupation, port traffic, hardware temperature, etc.) and service data (such as user concurrency, data throughput, service response time, etc.). Each original data point carries metadata information such as timestamp, device identifier, and service type, for example, a data record is: timestamp device utilization rate, these original data points are stored in the data buffer in chronological order.

[0160] Obtain the final health adjustment coefficient of each heterogeneous sub-region generated in step 403, which represents the health degree of the device state in the sub-region (value range 0-1, the higher the value, the more stable the state). According to the "point cloud data-sub-region" mapping relationship established in step 400, each original data point is associated with the corresponding heterogeneous sub-region. For example:

[0161] If a certain original data point corresponds to a point cloud located in sub-region A, the health adjustment coefficient of sub-region A is obtained =0.8; if it is located in sub-region B, the health adjustment coefficient of sub-region B is obtained =0.3. Finally, a corresponding relationship table of "original data point-health adjustment coefficient" is formed.

[0162] Step 501, for each original data point, adjust the coefficient according to the health degree of the heterogeneous sub-region it belongs to. The core logic of the weighting process is that the data points corresponding to the sub-regions with high health degree (coefficient close to 1) have higher weights, and the data points corresponding to the sub-regions with low health degree (coefficient close to 0) have lower weights. The specific steps are as follows:

[0163] Suppose a certain original data point D contains 3 indexes (CPU utilization, memory occupation, user concurrency), and the health degree adjustment coefficient of the sub-region it belongs to is R; multiply each index value in D by the coefficient R to obtain the weighted data point D'. For example: the original data point If the sub-region health degree coefficient is 0 (extreme abnormal state), then all index values of the corresponding data point are weighted to approach 0, reducing the interference of abnormal data on model training.

[0164] Through weighting, data during stable operation of the device is given higher weight, highlighting the characteristics of the normal state; data in abnormal or unstable state is given lower weight, reducing noise interference. For example, the influence of data points caused by device aging leading to performance fluctuations (corresponding to low health degree sub-region) is weakened in model training, while data during stable operation of the device will dominate the construction of the threshold model.

[0165] Step 502, the weighted data points still have dimension differences in different index dimensions (such as CPU utilization is percentage, user concurrency is integer), in order to unify the data scale and improve the model training effect, the system adopts the minimum-maximum normalization method to map each index value of all data points to the [0, 1] interval. The specific process is as follows:

[0166] For each index dimension (such as CPU utilization, memory occupation, etc.), find the minimum and maximum values of that dimension in the preprocessed data set;

[0167] For any data value e of that dimension, use the formula e' to calculate the normalized value e'.

[0168] Normalization makes indexes of different dimensions comparable, avoiding the dominance of indexes with large values (such as user concurrency) in model training, while accelerating the convergence speed of the model. For example, if not normalized, small changes in user concurrency (value may reach thousands) may mask significant fluctuations in CPU utilization (value range 0-100%), while after normalization, the influence weights of the two on the model are consistent.

[0169] Step 503, select historical data (such as data in the past week or month) from the preprocessed dataset as training input data. When dividing, the time sequence of the data is preserved to ensure that the model learns the time sequence characteristics of the data. For example, if the preprocessed dataset contains data from 2024-01-01 to 2024-01-31, select the data from the first 20 days (2024-01-01 to 2024-01-20) as the training set, and the last 10 days as the validation set. Use adaptive modeling algorithms (such as long short-term memory network LSTM, gradient boosting tree, etc.) to learn the training data, the specific process is as follows:

[0170] Automatically identify the correlation between device performance data and business data (such as the synchronous change trend of CPU utilization and user concurrency); by analyzing historical data, learn the performance change pattern of the device under different time periods (such as weekdays / weekends, daytime / nighttime), and different business loads; according to the feedback results of the validation set, adjust the model parameters (such as the weights of the neural network, the split threshold of the tree model) to avoid overfitting or underfitting. Finally, the model can capture the dynamic change pattern of the data, for example, predict the CPU utilization fluctuation range of the device during the business peak period.

[0171] Step 504, based on the training results, generate a dynamic threshold rule set, define dynamic threshold intervals for each device performance indicator and business indicator. The construction logic of the rule set is as follows:

[0172] Calculate the statistical distribution (such as mean, standard deviation, quantile) of each indicator value in the training data; determine the threshold boundary according to the statistical distribution, for example:

[0173] Normal threshold lower limit = mean - 2 x standard deviation;

[0174] Normal threshold upper limit = mean + 2 x standard deviation;

[0175] Combine time factors (such as weekday / weekend patterns), business types (such as voice / data services), etc. to dynamically correct the threshold interval. For example, the user concurrency threshold upper limit is adjusted lower during weekends when user concurrency is usually lower. Encapsulate the dynamic threshold rule set into an executable dynamic threshold model, which has the following functions:

[0176] Real-time data input: receive real-time preprocessed data (weighted and normalized device and business data);

[0177] Threshold matching judgment: automatically judge whether the input data is outside the threshold interval, output normal / abnormal identification;

[0178] Interface call: provide API interface, support integration with monitoring system, early warning module, realize real-time response of abnormal data.

[0179] By weighting through the health adjustment coefficient, the importance of the stable state data of the equipment is highlighted, the abnormal data interference is suppressed, the model training is more focused on the normal operation rules, and the accuracy of the threshold model is improved. The weighted normalization solves the problem of inconsistent dimensions of the original data, ensures the balanced contribution of each index to the model training, avoids the key information being covered, and enhances the adaptability of the model to complex data. The adaptive algorithm learns the dynamic change mode of the historical data, generates a threshold interval that changes with time and business load, replaces the traditional fixed threshold, and effectively reduces the false positive and false negative rates. The dynamic threshold model not only can detect abnormalities in real time, but also can predict the performance trend of the equipment based on historical rules, provide quantitative basis for predictive maintenance, and promote the operation and maintenance mode from passive response to active prevention.

[0180] In a preferred embodiment of the present application, step 6, the real-time preprocessed data is input into the dynamic threshold model for matching, a warning is triggered when the data deviates from the threshold, and the physical mapping relationship of the abnormal point aggregation sub-area is associated to locate the faulty equipment; based on the time sequence change trend of the health adjustment coefficient, the output result of the dynamic threshold model is combined to predict the equipment failure and generate a set of predictive maintenance instructions, which can include:

[0181] Step 600, input real-time preprocessed data into the dynamic threshold model for matching, and trigger an equipment abnormality warning signal when the data exceeds the dynamic threshold interval;

[0182] Step 601, extract the abnormal data points associated with the warning signal and the identification of the heterogeneous sub-area, locate the target faulty equipment based on the physical mapping relationship of the sub-area;

[0183] Step 602, continuously collect the health adjustment coefficients of each heterogeneous sub-area, generate a time sequence change sequence, analyze the change rate and fluctuation intensity characteristics of the time sequence, and input the change rate and fluctuation intensity characteristics into the dynamic threshold model to output the equipment failure probability;

[0184] Step 603, when the failure probability exceeds a preset threshold, a set of predictive maintenance instructions is generated.

[0185] In an embodiment of the present application, the weighted and normalized data processed in step 502 is obtained from real-time data stream, each data point containing device operation indicators (such as CPU utilization, memory occupancy, network traffic) and business indicators (such as user concurrency, response time), and has been converted into dimensionless values in the interval [0, 1]. The data point carries a timestamp, a device identifier, and an identifier of the heterogeneous sub-region to which it belongs, for example: [timestamp: 2024-02-15 14:30:00, device ID: DEV012, sub-region ID: S03, CPU utilization: 0.8, memory occupancy: 0.7, user concurrency: 0.6] The data point is input into the dynamic threshold model, which automatically identifies the indicator type (such as CPU utilization corresponding to device performance indicators) corresponding to the data. The dynamic threshold model pre-stores a dynamic threshold interval for each indicator based on historical data training, for example:

[0186] CPU utilization threshold interval: [0.2, 0.75] (weekday daytime period);

[0187] User concurrency threshold interval: [0.1, 0.5] (current business type).

[0188] Compare the indicator value of the input data with the corresponding threshold interval. If a certain indicator value (such as CPU utilization = 0.8) exceeds the upper threshold (0.75) or is lower than the lower threshold (0.2), the indicator is determined to be abnormal.

[0189] When one or more indicators of a single device are abnormal, a device abnormality early warning signal is triggered immediately, and the abnormal time, abnormal indicators, and over-standard values are recorded.

[0190] Step 601, after receiving the early warning signal, extract the abnormal data point that triggered the early warning and the identifier of the heterogeneous sub-region to which it belongs (such as sub-region ID: S03). Each heterogeneous sub-region has established a mapping relationship with the actual device in the grid division stage of step 301, for example:

[0191] Sub-region S03 contains devices DEV012, DEV015, and DEV018;

[0192] Sub-region S05 contains devices DEV020 and DEV021.

[0193] Combine the device ID of the abnormal data point with the sub-area mapping relationship to determine the scope of the faulty device. If the device ID of the abnormal data point is DEV012 and the sub-area ID is S03, then directly locate DEV012 as the faulty device. If multiple devices in the sub-area all have abnormalities (e.g., DEV012 and DEV015 in S03 exceed the standard at the same time), then further analyze the similarity of the abnormal indicators (e.g., whether they are all CPU utilization abnormalities) to narrow down the fault range to the device group with common abnormalities (e.g., judge that the cooling system in the S03 area affects multiple devices).

[0194] Step 602: Continuously collect the health adjustment coefficients of each heterogeneous sub-area (generated in step 403) and arrange them in chronological order to form a time series. For example, the health coefficient sequence of sub-area S03 is: [2024-02-01:0.8, 2024-02-02:0.78, 2024-02-03:0.75,..., 2024-02-15:0.6]. This sequence reflects the long-term trend of device status in the sub-area. The health coefficient sequence is analyzed as follows:

[0195] Calculate the coefficient change rate of adjacent time points, for example, the change rate from 2024-02-02 to 2024-02-03 is ≈-0.03846154, negative value indicates health decline;

[0196] Measure the fluctuation degree by calculating the standard deviation or variance of the sequence. The greater the fluctuation intensity, the more unstable the device status;

[0197] Use moving average or polynomial fitting method to predict the future trend of health coefficient (such as judging whether it continues to decline).

[0198] Combine the change rate, fluctuation intensity and other features with the dynamic threshold model to output the device failure probability:

[0199] The model is trained based on historical data to establish a mapping relationship between "health features and failure probability" (for example, when the health coefficient decline rate is greater than 0.05 and the fluctuation intensity is greater than the threshold, the failure probability significantly increases); input the current health features, the model calculates the failure probability of the device in the future (such as the next 24 hours) through regression algorithm (such as logistic regression), the value range is 0-100%.

[0200] Step 603: Pre-set failure probability threshold (such as 80%), when the failure probability output by the dynamic threshold model exceeds the threshold, it is determined that the device has high risk failure. For example, if the failure probability of device DEV012 is calculated as 85%, the predictive maintenance process is triggered. According to the failure probability, abnormal indicator type and device historical maintenance record, a targeted maintenance instruction set is generated:

[0201] Fault type localization: combined with abnormal indicators (such as CPU overheating, memory leakage) and health degree decline trend, infer potential fault causes (such as cooling fan failure, software memory overflow);

[0202] Maintenance operation suggestion: specific maintenance steps, for example:

[0203] If it is determined that the heat dissipation problem, the instruction set includes "check the cooling fan speed" "clean the dust in the device heat dissipation hole";

[0204] If it is a software failure, the instruction set includes "restart the service process" "update to the latest version";

[0205] Priority marking: according to the severity of the fault, mark the maintenance priority (such as emergency, high, medium, low), emergency fault (such as power module failure risk) needs to be handled immediately, and low priority fault can be arranged in non-peak period.

[0206] Push the maintenance instruction set to the terminal of the operation and maintenance personnel (such as mobile phone APP, monitoring platform), and associate the device geographic location, maintenance manual link and other information, assist the operation and maintenance personnel to quickly execute the maintenance task. At the same time, the system records the instruction execution state, and closes the loop management maintenance process.

[0207] The dynamic threshold model dynamically adjusts the threshold based on historical data, which reduces the false alarm and missed alarm rate compared with the fixed threshold, realizes the second-level abnormal response combined with real-time data matching, and guarantees the system stability. Through the mapping relationship between heterogeneous sub-regions and devices, the fault localization is reduced from the fuzzy system level to specific devices or device groups, reducing the troubleshooting time and improving the fault handling efficiency. Based on the time series analysis of the health degree coefficient and the dynamic model, the performance degradation trend of the device is identified in advance, the operation and maintenance is changed from "post-fault repair" to "pre-fault prevention", and the risk of sudden shutdown is reduced. The predictive maintenance instruction set generates customized operation suggestions according to the fault probability and device characteristics, avoids excessive maintenance or insufficient maintenance, optimizes the allocation of operation and maintenance resources, and reduces the maintenance cost. From the whole process automation of abnormal early warning, fault localization to maintenance execution, realize the digitalization and intelligentization upgrade of operation and maintenance management, improve the overall reliability and operation and maintenance efficiency of the communication system.

[0208] As shown in Figure 2 The embodiment of the application also provides a communication system operation and maintenance management system applying data analysis, which comprises:

[0209] A point cloud generation module is configured to collect raw data streams of device operation data and business data in real time, divide the raw data streams according to a preset time window, and map multi-dimensional data in each window into a point cloud data set in a high-dimensional space;

[0210] A plane structure module is configured to extract spatial distribution features by principal component analysis based on a point cloud dataset, and construct a virtual reference plane representing a state;

[0211] A sub-region generation module is configured to demarcate an initial quadrilateral mesh on the virtual reference plane, and perform adaptive segmentation on the initial mesh according to a density gradient of the point cloud dataset to form heterogeneous sub-regions;

[0212] A health degree calculation module is configured to map the point cloud dataset to the heterogeneous sub-regions, count distribution entropy values of point clouds in each sub-region, and calculate a health degree adjustment coefficient of each sub-region based on the distribution entropy values;

[0213] A model training module is configured to perform weighted and normalized processing on the original data and the health degree adjustment coefficient to generate a preprocessed dataset, and train a dynamic threshold model of device performance and business data based on the preprocessed dataset by an adaptive modeling algorithm;

[0214] A maintenance instruction generation module is configured to input real-time preprocessed data into the dynamic threshold model for matching, trigger an early warning when the data deviates from a threshold, locate a faulty device by associating a physical mapping relationship of an abnormal point cluster, and predict a device fault based on a time sequence change trend of the health degree adjustment coefficient and an output result of the dynamic threshold model, and generate a set of predictive maintenance instructions.

[0215] It should be noted that the system corresponds to the above method, and all implementation manners in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0216] Embodiments of the present application also provide a computing device, comprising a processor and a memory storing a computer program, wherein the computer program is executed by the processor to perform the above method. All implementation manners in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0217] Embodiments of the present application also provide a computer readable storage medium storing instructions, wherein the instructions are executed on a computer to make the computer perform the above method. All implementation manners in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0218] The above is the preferred embodiment of the present application. It should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should also be considered within the scope of protection of the present application.

Claims

1. A communication system operation and maintenance management method based on data analysis, characterized in that, The method includes: Step 1: Collect raw data streams of equipment operation data and business data in real time, divide the raw data streams into preset time windows, and map the multidimensional data in each window into point cloud datasets in high-dimensional space; Step 2: Perform principal component analysis on the point cloud dataset to calculate the eigenvectors and corresponding eigenvalues ​​of the covariance matrix; filter the eigenvectors based on the magnitude of the eigenvalues ​​to determine the basis vectors used to characterize the main spatial distribution features; span a two-dimensional plane from the basis vectors to serve as a virtual reference plane characterizing the spatial distribution features of the point cloud dataset. Step 3: Define an initial quadrilateral mesh on the virtual reference plane, and adaptively segment the initial mesh according to the density gradient of the point cloud dataset to form heterogeneous sub-regions; Step 4: Map the point cloud dataset to heterogeneous sub-regions, calculate the distribution entropy value of the point cloud in each sub-region, and calculate the health adjustment coefficient of each sub-region based on the distribution entropy value. Step 5: Perform weighted normalization on the original data and health adjustment coefficients to generate a preprocessed dataset. Based on the preprocessed dataset, train an adaptive modeling algorithm to generate a dynamic threshold model for device performance and business data. Step 6: Input the real-time preprocessed data into the dynamic threshold model for matching. When the data deviates from the threshold, an early warning is triggered, and the physical mapping relationship of the sub-regions where abnormal points are clustered is associated to locate the faulty equipment. Based on the time-series change trend of the health adjustment coefficient and combined with the output results of the dynamic threshold model, the equipment failure is predicted, and a predictive maintenance instruction set is generated.

2. The communication system operation and maintenance management method based on application data analysis according to claim 1, characterized in that, The system collects raw data streams of equipment operation data and business data in real time, divides the raw data streams into preset time windows, and maps the multidimensional data within each window into a point cloud dataset in a high-dimensional space, including: The collected raw data stream is divided into time windows according to a preset fixed duration; For each time window, an n-dimensional data vector is converted into coordinate points in an m-dimensional space; Aggregate all the coordinate points in the m-dimensional space obtained from the transformation within each time window to form a point cloud dataset representing the state of that time window.

3. The communication system operation and maintenance management method based on application data analysis according to claim 2, characterized in that, An initial quadrilateral mesh is defined on a virtual reference plane, and the initial mesh is adaptively segmented based on the density gradient of the point cloud dataset to form heterogeneous sub-regions, including: Define an initial quadrilateral grid on the virtual reference plane and calculate the density distribution of the projected points of the point cloud dataset on the virtual reference plane. Based on the density distribution of the projection points, the density gradient of each part within the initial quadrilateral grid region is determined, and an adaptive region segmentation operation is performed on the initial quadrilateral grid according to the density gradient. Through adaptive region segmentation, the virtual reference plane is divided into grid regions with different point cloud density characteristics. These grid regions are called heterogeneous sub-regions.

4. The communication system operation and maintenance management method based on application data analysis according to claim 3, characterized in that, The point cloud dataset is mapped to heterogeneous sub-regions, the distribution entropy value of the point cloud in each sub-region is calculated, and the health adjustment coefficient of each sub-region is calculated based on the distribution entropy value, including: The points in the point cloud dataset are mapped to the divided grid regions on the virtual reference plane, so that each point uniquely falls into a corresponding heterogeneous sub-region, forming a point cloud subset corresponding to each heterogeneous sub-region. For each heterogeneous subregion, the distribution entropy value of the corresponding point cloud subset is calculated based on the spatial distribution characteristics of the corresponding point cloud subset within the corresponding heterogeneous subregion. Based on the distribution entropy value of each heterogeneous sub-region, the initial health weight factor for each heterogeneous sub-region is determined through a preset entropy-health mapping relationship. The initial health weighting factors of each heterogeneous sub-region are normalized to generate the final health adjustment coefficient for each heterogeneous sub-region.

5. The communication system operation and maintenance management method based on application data analysis according to claim 4, characterized in that, The raw data and health adjustment coefficients are weighted and normalized to generate a preprocessed dataset. Based on this preprocessed dataset, a dynamic threshold model for device performance and business data is trained using an adaptive modeling algorithm, including: Obtain the original data points in the original data stream and the final health adjustment coefficient for each heterogeneous sub-region; Based on the heterogeneous sub-region to which each original data point belongs, the corresponding final health adjustment coefficient is used to weight the original data points to generate weighted data points. All weighted data points are normalized to form a preprocessed dataset; The historical data portion of the preprocessed dataset is used as training input data. An adaptive modeling algorithm is used to learn the dynamic change patterns of device performance data and business data in the training input data to obtain the learning results. Based on the learning results, a dynamic threshold rule set is generated. The rule set defines the dynamic threshold range of device performance and business data, and the dynamic threshold rule set is encapsulated into a dynamic threshold model.

6. The communication system operation and maintenance management method based on application data analysis according to claim 5, characterized in that, Real-time preprocessed data is input into a dynamic threshold model for matching. When the data deviates from the threshold, an early warning is triggered, and the physical mapping relationship of sub-regions where anomalies are clustered is associated to locate faulty equipment. Based on the time-series variation trend of the health adjustment coefficient and combined with the output of the dynamic threshold model, equipment failures are predicted, and a predictive maintenance instruction set is generated, including: Real-time preprocessed data is input into a dynamic threshold model for matching. When the data exceeds the dynamic threshold range, an abnormal warning signal for the device is triggered. Extract abnormal data points associated with the early warning signal and their corresponding heterogeneous sub-region identifiers, and locate the target faulty equipment based on the physical mapping relationship of the sub-regions; The health adjustment coefficients of each heterogeneous sub-region are continuously collected to generate a time-series change sequence. The change rate and fluctuation intensity characteristics of the time-series sequence are analyzed, and the change rate and fluctuation intensity characteristics are input into a dynamic threshold model to output the probability of equipment failure. When the probability of failure exceeds a preset threshold, a predictive maintenance instruction set is generated.

7. A communication system operation and maintenance management system that applies data analysis, wherein the system implements the method as described in any one of claims 1 to 6, characterized in that, include: The point cloud generation module is used to collect raw data streams of equipment operation data and business data in real time, divide the raw data streams according to preset time windows, and map the multi-dimensional data in each window into point cloud datasets in high-dimensional space. The planar construction module is used to perform principal component analysis on point cloud datasets and calculate the eigenvectors and corresponding eigenvalues ​​of the covariance matrix. Feature vectors are selected based on the magnitude of their eigenvalues ​​to determine the basis vectors used to characterize the main spatial distribution features; A two-dimensional plane is spanned by the basis vectors, serving as a virtual reference plane characterizing the spatial distribution features of the point cloud dataset; The sub-region generation module is used to define an initial quadrilateral mesh on a virtual reference plane and adaptively segment the initial mesh according to the density gradient of the point cloud dataset to form heterogeneous sub-regions. The health calculation module is used to map the point cloud dataset to heterogeneous sub-regions, calculate the distribution entropy value of the point cloud in each sub-region, and calculate the health adjustment coefficient of each sub-region based on the distribution entropy value. The model training module is used to perform weighted normalization on the raw data and health adjustment coefficients to generate a preprocessed dataset, and based on the preprocessed dataset, to train and generate a dynamic threshold model for device performance and business data through an adaptive modeling algorithm. The maintenance instruction generation module is used to input real-time preprocessed data into a dynamic threshold model for matching. When the data deviates from the threshold, an early warning is triggered, and the physical mapping relationship of sub-regions where anomalies are clustered is associated to locate faulty equipment. At the same time, based on the time-series change trend of the health adjustment coefficient and combined with the output results of the dynamic threshold model, the module predicts equipment failures and generates a predictive maintenance instruction set.

8. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images